Earlier this month, Nvidia CEO Jensen Huang walked onto a stage in San Jose and declared that every enterprise CEO on earth now needs an OpenClaw (an open-source AI agent) strategy. Not just a considered response to it. Not a plan to evaluate it. A strategy. Already. He compared it to the moment every company needed a Linux strategy, an HTTP strategy, a Kubernetes strategy. Those were the architectural shifts that defined their eras. Jensen was saying this is the next one, and that it had arrived faster than any of those before.
What Huang announced at GTC is not a single product or a single bet; it is a unified play through a series of interlocking moves that describe a company engineering the conditions for the next phase of computing. In so doing, the company positions itself as the indispensable infrastructure layer beneath it.
The unified play
The centrepiece of the GTC keynote was the announcement that Nvidia were moving to commercial deployment of Groq, its US$20 billion licensing bet on the inference era.
While AI training tolerates high latency, inference is different. Inference is the act of answering: a chatbot response, an agent executing a task, or a model generating output in real time. It is latency-sensitive, persistent, and increasingly the dominant computational workload as AI moves from the laboratory into daily commercial use.
By 2030, it is estimated that seventy per cent of all data centre demand will originate from inference. The economics of how that inference is delivered will define the economics of the AI era.
For Nvidia, the problem is that the GPU is extraordinary at training but is architecturally mismatched for inference decode. At the low batch sizes typical of interactive inference, the GPU’s arithmetic units sit idle, waiting for memory. The model must load its weight parameters from memory for every token generated, which, for a large model, is hundreds of gigabytes traversing the memory bus on every forward pass. This is where efficiency is lost and costs accumulate.
Groq’s architecture solves this by inverting the conventional design. Where a GPU stores weights in external high-bandwidth memory and fetches them at runtime, Groq’s Language Processing Unit places weights directly in on-chip SRAM, achieving significant memory bandwidth of between 20 and 50 times more throughput than a GPU. In Groq’s LPU, the compiler, not the hardware, schedules every operation in advance. Nothing waits. Nothing speculates. The result is thirty-five times more tokens per second per megawatt than Blackwell alone. That is a genuine architectural advantage.
The Vera Rubin architecture integrates Groq’s LPU alongside the GPU in a disaggregated inference design: Rubin handles the compute-intensive prefill phase, Groq handles the memory-bandwidth-bound decode phase. These two tasks, now on different hardware, require a high-bandwidth, low-latency optical link to pass computation between them.
Nvidia has this covered and recently took concurrent investments of approximately US$4 billion across both Coherent and Lumentum as the two dominant vertically integrated optical component suppliers. They are the infrastructure layer required to make disaggregated inference work at scale. This is important given it brings another technology transition to the fore.
Copper versus optical. The shift from copper to optical interconnects is not a footnote in disaggregated inference, and AMD, Broadcom, Nvidia, OpenAI, Meta, and Microsoft have agreed to collaborate through the Optical Compute Interconnect Multi-Source Agreement (OCI MSA) group. This cohort is tasked with defining an open connectivity specification for optical interconnections which the ongoing AI infrastructure buildout demands.
The business logic running beneath all of this is transparent. The agent era is here; use cases are exploding and the growth in inference requirements is growing exponentially. For 2026, the hyperscalers have increased their outlook for AI-related capital expenditure significantly, with 75% expected to be on AI-related hardware. Inference is in short supply.
OpenClaw has set alight the AI inference touch paper and is responsible for launching an agent framework that is both free and open – and is potentially where all the economic value accrues to the hardware and infrastructure on which those agents run.
Nvidia’s Nemotron models, running locally on Nvidia hardware, remove the token cost for the agent’s underlying intelligence entirely. Nvidia has announced NemoClaw, which, although in its infancy, represents the company’s enterprise-hardened wrapper that has security controls and policy-based governance. This potentially removes an adoption barrier, and it is expected to install in a single command.
Every agent that runs, every token consumed, every enterprise that builds an OpenClaw strategy will drive compute demand back to Nvidia’s silicon. The catalyst that lit this fuse arrived from an unexpected direction.
The black swan
OpenClaw itself began life in November 2025 as Clawdbot, before briefly being renamed Moltbot after Anthropic raised trademark concerns. Within weeks, OpenClaw became the fastest-growing open-source project in the history of computing. This release showed what was possible and where traditional sensibilities were too afraid to tread.
Someone I know described OpenClaw as a black swan event, a phenomenon described and brought into popular culture by Nassim Nicholas Taleb in his 2007 book of the same name. Such an event has three attributes: rarity, extremeness of impact, and retrospective predictability.
For the very few, the arrival of autonomous AI agents capable of writing, reasoning, coding, and operating software will not have been a surprise, but for the enterprise software industry and its investors, it is. OpenClaw is not a better version of what came before. It replaces our foundational understanding of how we interact and solve with technology.
In retrospect, OpenClaw is characteristic of how these moments arrive – obliquely, without announcement and from a direction nobody was watching.
The front door
The technology shift needed a vehicle to reach the enterprise world at speed. OpenClaw provided it. And in doing so, it forced into the open a question that every software business had been quietly avoiding: who owns the interface between the user and the software they use?
Sridhar Ramaswamy, formerly the architect of Google’s advertising business and now the CEO of Snowflake, framed it at the recent Morgan Stanley conference: Google’s first rule was always to own on the front door.
A business giving Google the search function believing that it would retain the relationship was, to put it mildly, a risk that usually led to the discovery that Google had rummaged through their pockets and left them empty. Today, the leading AI operators are positioning to become the interface layer through which users interact with every software environment they use. When an AI agent can seamlessly operate a CRM or financial modelling tool, the relationship shifts and the plumbing, not the software vendor, becomes the relationship.
This is not a hypothetical. Late last year, Anthropic released Claude for Excel and PowerPoint, establishing a direct relationship with enterprise users at ‘the front door’ – from within Microsoft’s own applications, without the need for a Microsoft AI license.
Was this a Trojan Horse? Now, a few short months later, Microsoft has answered the question itself: it announced the integration of Anthropic’s Claude into Copilot across Outlook, Teams, and Excel, representing the most consequential instance of this dynamic at scale. The standalone Claude add-ins for Excel and PowerPoint are Anthropic’s own beachhead; the Copilot integration is Microsoft’s response to it. Microsoft controls roughly 450 million enterprise software seats and they have had to cede this ground. Copilot has failed so far – has Microsoft handed the keys of its front door to the most capable AI operator in the market?
Looking at it critically from the outside in, it is easy to conclude there are non-trivial concerns within Microsoft. Nadella, the CEO, is stationed at the front line of engineering and is personally overseeing Copilot. He is a focused executioner and will be marshalling all the company’s significant resources, but the clock is ticking. The longer the cojoining of Microsoft and Anthropic endures, the greater the unease investors will feel. That Nadella is at the front line may be feel reassuring, but perhaps this signals how dire the situation really is. Perhaps Anthropic is already putting on its slippers and making itself comfortable by the fire.
The redundancy storm
What arrives immediately behind the front door disruption is a consequence that investors have treated as gradual and distant. It is neither. Huang himself made the arithmetic explicit: a highly paid engineer not consuming AI tokens at roughly half their annual salary is not, in his framing, genuinely productive. The mathematics are direct – you can maintain the same total cost base with fewer people.
The mechanism is worth understanding. AI has broken the informational advantage held by engineers who originally built a system – the knowledge of where the complexity was buried, the architecture that took years to learn. That knowledge was job security. It no longer is. An AI can reverse-engineer an existing code base in an afternoon, enabling a less experienced developer anywhere in the world to take over at a fraction of the cost.
This is not a future risk. The displacement is already visible in the data. Technology sector layoffs have accelerated in 2026, with the effects disproportionately targeted at mid-to-senior engineering roles and not the junior staff AI was first expected to replace; however, graduate hiring in software engineering has also fallen sharply. The calculus of training humans against deploying agents has shifted.
The ripple does not stop at the technology sector’s boundary. The professional middle class that has been the engine of discretionary consumer spending is seeing consolidation, and unless they are able to retrain or deploy their skills elsewhere, the pace of long-term consumer spending may be revised lower. Uncertainty leads to modification of spending habits. The macroeconomic transmission mechanism from AI-driven may be significant as consumers rein back their discretionary spending. The storm does not stay inside the data centre.
The honest reckoning
The investment implication follows directly, even if it is uncomfortable to state plainly. When a single open-source release can render years of accumulated advantage obsolete overnight – when the informed consensus about which businesses are defensible can shift faster than a capital allocation cycle – the terminal value assumptions embedded across a wide range of asset classes become genuinely difficult to ascertain.
The difficulty is not simply that forecasts need revising. It is that the analytical frameworks used to build those forecasts were themselves constructed in a world where competitive moats deepened slowly, where switching costs were structural and durable, and where the knowledge embedded in a workforce was an asset that compounded over time. Each of those assumptions is now under threat. Software businesses that priced years of enterprise stickiness into their multiples face a different question: not whether AI will affect them, but how quickly the interface layer above them gets claimed by someone else. The front door problem is a valuation problem.
In that context, perhaps the relative value of businesses anchored in the physical world deserves a fresh look; not because they have become more exciting, but because they have become comparatively rarer. The physical world does not version-update. A mine requires permits, capital, and geological reality. Industrial infrastructure ages on predictable timelines. Energy assets are subject to physics, not to the next model release. These are not exciting characteristics. In a market that has spent several years re-rating excitement, they may be exactly the point.
This is not a call to retreat from technology. The infrastructure layer beneath the agent era – the silicon, the optical interconnects, the power – remains among the most compelling capital deployment opportunities in a generation. But the middle ground, the application software businesses that sit above the infrastructure and below the intelligence, is where the pressure has coalesced. Disruption will be widespread, and it is where valuation assumptions deserve the most scrutiny and where the distance between current prices and honest reckonings may be the widest.
The black swan has arrived. It has paddled to the shore.
Tim Chesterfield is CIO of the Perpetual Guardian Group and the founding CIO and Director of its investment management business, PG Investments. With $2.8 billion in funds under management and $8 billion in total assets under management, Perpetual Guardian Group is a leading financial services provider to New Zealanders.
Disclaimer
Information provided in this publication is not personalised and does not take into account the particular financial situation, needs or goals of any person. Professional investment advice should be taken before making an investment. The information provided in this article is not a recommendation to buy, sell, or hold any of the companies mentioned. PG Investments is not responsible for, and expressly disclaims all liability for, damages of any kind arising out of use, reference to, or reliance on any information contained within this article, and no guarantee is given that the information provided in this article is correct, complete, and up to date.
